Separating Expert Retention from Autonomous Source Inference in Raw-ECG-Replay-Free Continual ECG Deployment

What Happened

This research from arXiv (2607.01674) addresses a critical deployment challenge in medical AI: how to maintain performance when a multi-source ECG model encounters new data sources after the original raw training data can no longer be retained or replayed. The authors propose decoupling two distinct capabilities—expert retention (preserving knowledge of previously seen sources) and autonomous source inference (identifying which source new data belongs to without explicit labels).

The core technical approach involves freezing a pretrained backbone network and assigning each data source its own isolated classifier head. This prevents parameter interference between sources, a common failure mode in continual learning where updating weights for new data degrades performance on older sources. By keeping the backbone fixed and only adding new classifier heads for unseen sources, the model avoids catastrophic forgetting without requiring access to the original raw ECGs.

Why It Matters

This work tackles a real-world constraint that is often ignored in academic continual learning benchmarks: privacy and regulatory compliance. In clinical settings, raw ECG data cannot always be stored indefinitely or replayed during retraining due to patient consent limitations, data protection laws like GDPR, or hospital data retention policies. Most existing continual learning methods assume some form of data replay or rehearsal buffer, which is impractical here.

The separation of expert retention from source inference is particularly elegant. In multi-source deployment, the model must not only remember how to classify ECGs from previously seen hospitals or devices (expert retention), but also recognize when a new source arrives without being explicitly told (autonomous source inference). By isolating classifiers per source, the model can add new experts without disrupting old ones, and the frozen backbone provides a stable feature space for comparing incoming data against known source distributions.

Implications for AI Practitioners

For engineers deploying medical AI systems, this research offers a practical blueprint for handling data-source drift without costly retraining cycles. The approach is lightweight—adding a classifier head is far cheaper than fine-tuning an entire backbone—and aligns with real-world constraints where raw data access is temporary.

However, practitioners should note that this method assumes the pretrained backbone is sufficiently general to represent all future sources. If a new source introduces fundamentally different ECG morphologies (e.g., pediatric vs. adult populations), the frozen backbone may lack the capacity to encode those features. The paper does not address how to detect or handle such distributional shifts that require backbone adaptation.

Additionally, the autonomous source inference component likely requires some form of uncertainty estimation or outlier detection to trigger the creation of a new classifier head. Practitioners will need to implement a thresholding mechanism that balances sensitivity (catching true new sources) against false positives (creating unnecessary heads for noise).

Key Takeaways

• Decoupling retention from inference enables continual ECG deployment without raw data replay, solving a real regulatory and privacy constraint in clinical AI.
• Frozen backbone + per-source classifiers prevents catastrophic forgetting while allowing the model to scale to new data sources, but assumes the backbone generalizes to all future source distributions.
• Autonomous source detection is a necessary but underexplored component—practitioners must implement robust outlier detection to trigger new classifier creation without human intervention.
• Lightweight deployment makes this approach attractive for resource-constrained clinical environments, but validation on diverse ECG populations is needed before production use.